Breaking

OpenAI Unveils GPT-Red, an Automated Red Teamer to Hunt Prompt-Injection Flaws

July 15, 2026 at 13:35 EDT

Security
Foundation Models
AI Agents

OpenAI has introduced GPT-Red, an internal, automated red teamer designed to uncover prompt-injection vulnerabilities in its models at scale and harden defenses before wider deployment. The tool was detailed in a company blog post, GPT-Red: Unlocking Self-Improvement for Robustness, which frames the system as a way to use today's models to make tomorrow's models more resistant to attack.

OpenAI · GPT-Red

An AI that attacks AI — to make it harder to hack

GPT-Red is OpenAI's internal, automated red teamer. Trained by self-play, it hunts prompt-injection flaws at scale and feeds every successful attack back into the defender — hardening production models before they ship.

6×

lower injection failure rate vs. the best model from 4 months earlier

0.05%

direct prompt-injection failure rate for GPT-5.6 Sol

large post-training run of compute to train it

Indirect-injection arena · attack success rate

Machine attacker vs. human attacker

84%

GPT-Red

13%

Human red teamers

Automated probing surfaces roughly 6× more working attacks than manual testing at scale.

The self-improvement loop

Attacker

crafts prompt injections

→

Defender

tries to resist the attack

→

Reinforce

wins train a stronger defender

Repeat via self-play reinforcement learning — a scalable safety flywheel.

Proven exploits — kept internal only

Compromised "Vendy," an in-office AI vending machine — altered prices and canceled orders.
Ran a data-exfiltration attack on a Codex CLI agent, more effective and token-efficient than baselines.

The offensive capability is never released — it only injects robustness into shipping models.

Positive reception

Developers call the self-play design a scalable safety flywheel, and see the 6× robustness gain as a meaningful step against real injection attacks.

Open questions

How it meshes with external red-teaming, whether it generalizes beyond prompt injection, and independent validation — pending a forthcoming pre-print.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.